[Privacy Concern]: Agent mode is hard-coded to Claude Code SDK — no user opt-out, no fallback for OpenAI-compatible models

[ninjas-W]

Problem Statement

Cherry Studio's Agent feature is hard-coded to use @anthropic-ai/claude-agent-sdk as the sole execution engine. This creates several issues:

1. No support for OpenAI-compatible models: Users who configure OpenAI-compatible providers (like custom self-hosted models) cannot use Agent mode, even though these models work perfectly in Chat mode. The error message is: Claude Code returned an error result: There's an issue with the selected model...

2. Privacy concerns: Claude Code is a closed-source commercial SDK that may collect user data. Users have no visibility into what data is being sent to Anthropic's servers, and no option to disable this telemetry.

3. No user choice: The Agent system does not allow users to opt-out of Claude Code or choose an alternative Agent engine. There's no "local-only" or "OpenAI-compatible" mode.

4. No graceful degradation: When Claude Code is unavailable (no API key, network issues, or model incompatibility), Agent mode completely fails instead of falling back to a simpler but functional alternative.

Context

I understand that:

• Issue [Feature]: Suggest adjusting Agent framework to support OpenAI-compatible endpoints #14480, [Feature Request] Agent Function Support for OpenAI Compatible Endpoint Function Calling #14861, [Feature]: OpenAI / Anthropic API Format Interoperability #14873, and DeepSeek V4 from OpenCode Go package cannot be used in Agent mode #14918 have reported similar limitations
• PR feat(agents): add Claude OpenAI proxy for OpenAI-compatible providers #15177 added an internal proxy for CherryIN to support OpenAI-compatible providers through Claude Code
• The v2.0.0 milestone ([Feature]: LM Studio already supports Anthropic endpoint, but Agent creation still cannot use LMS models #12980) plans to refactor the Agent architecture

However, these solutions all assume Claude Code remains the mandatory execution engine — they just add translation layers. This doesn't address:

• The fundamental lack of user choice
• Privacy concerns about sending all Agent traffic through a closed-source SDK
• The desire for a truly provider-agnostic Agent framework

Desired Solution

I would like Cherry Studio to:

1. Make Claude Code optional: Allow users to explicitly choose whether to enable Claude Code-based Agent mode. If disabled, Agent functionality should gracefully degrade or use an alternative engine.

2. Support OpenAI-compatible Agent engines: Either:

   • Use a provider-agnostic Agent framework (like ai-sdk's agent patterns) that supports both Anthropic and OpenAI tool-calling protocols
   • Or provide a clear abstraction layer where users can plug in different Agent runtimes

3. Be transparent about data flow: Document what data Claude Code sends to external servers, and allow privacy-conscious users to opt-out.

4. Provide a local/privacy-first fallback: Even if limited, a local Agent engine (perhaps using simple ReAct patterns with any model) would be preferable to no Agent functionality at all for users who cannot or don't want to use Claude Code.

Alternative Workarounds

Currently, the only workaround is:

• Use Chat mode instead of Agent mode (loses tool orchestration, planning, and multi-turn capabilities)
• Run an external proxy like cc-switch or cc-nim (fragile, adds complexity, not user-friendly)

Additional Context

This is not just a "feature request" — it's about user agency and privacy. Tightly coupling Agent mode to a single closed-source SDK:

• Excludes users who cannot access Anthropic's API (geographic, financial, or compliance reasons)
• Creates a single point of failure
• Prevents community contributions to the Agent runtime
• Goes against the open-source spirit of Cherry Studio

I'm happy to contribute to a more modular Agent architecture if the maintainers are open to this direction.

[DeJeune]

Thanks for the detailed and thoughtful write-up.

A couple of updates on where things stand in v2:

• OpenAI-compatible models now work with Agent mode. In v2 the agent runtime routes non-Anthropic providers through a local translation gateway (Anthropic ⇄ OpenAI/other dialects), so an OpenAI-compatible provider can back an agent without you needing an external proxy like cc-switch/cc-nim. The old Claude Code returned an error result... path for incompatible models has been reworked.
• A runtime-driver abstraction is now in place, and additional agent engines (e.g. a pi-based runtime and an ai-sdk-based one) are actively in development. The goal is exactly the provider-agnostic, pluggable Agent runtime you describe — Claude Code becomes one driver among several rather than the mandatory engine.

So the pluggable/opt-out direction you asked for is the direction we are heading; the OpenAI-compatible part is already usable today in v2. We will keep this issue in mind as the additional engines land.

[ninjas-W]

This comment was translated automatically.

Thanks, looking forward to the V2 version

---

Original Content

感谢，很期待V2版本

[github-actions]

This issue has been inactive for a prolonged period and will be closed automatically in 10 days.
该问题已长时间处于闲置状态，10 天后将自动关闭。